Reflection began with bets on RL and Western open models

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 32:29

   I have other questions about the business, but before we get into that, for people who don't know, what's the origin story of Reflection? You guys work together at DeepMind. I actually heard that you reported to Yanis at one point. Yeah, Yanis was basically my boss. And now you're the CEO, so that's a fun twist. Well, we're actually equal partners. Your co-partners. That was going to be my question.

2. 32:49

   How do you run the company? You started it together, leaving DeepMind, maybe there first. Like, you bet on Open in, what, 2024? Is that right? Well, I guess it was a year ago. So, you know, we, well, actually, when we started the company, we had two bets. And I would kind of bring it back to the origin story. At the time, Giannis and I had worked on Gemini 1, 1.5. Giannis was the RLHF lead and, you know, one of the most storied reinforcement learning researchers at DeepMind.

3. 33:19

   I actually switched from physics to AI when I saw AlphaGo come out, and then I later learned, you know, well, I knew that Giannis was an author, but that was kind of a, you know, sort of a, dare I say it, a celebrity crush moment when I joined DeepMind and had the pleasure to work under Giannis on his RL team. But at the time we shipped, you know, Gemini 1 1.5, and those models were really good at chat.

4. 33:47

   You have to remember when ChatGPT came out, these models were not good at coding. They were not good at agentic. They were really good at chat. And we started the company with kind of two bets in mind, a scientific bet and kind of a market bet. One was that there would be, that reinforcement learning was the way to get these models to have agentic encoding capabilities. And the other one was that there'd be powerful Western open models we can build on top of so that we can just do this research a lot more cost effectively.

5. 34:20

   And the reinforcement learning within the first kind of year of the company became right across the industry. We were able to show some really strong reinforcement learning results at small scale. And obviously, the reasoning models started coming out at that time from other labs. But at the same time, the Western kind of open frontier receded with the DeepSeq v3 moment. And we actually waited around for a few months for something to come out that was as good and then saw it that wasn't happening. And so that's kind of how the company changed course. But there's, I think, even a better story in terms of how we actually started it. And when we were building Gemini, there would be these sprints where teams from